Plant material

Three pearl millet germplasm sets maintained in the Senegalese national collection (ISRA) were used, each evaluated in a separate trial; detailed pedigrees are provided in Supporting Information. The first set was a diversity panel of 290 genotypes from multiple breeding sources: 63 PMiGAP inbred lines, a reference panel from ICRISAT [12], 158 F5 ISMI (ISRA MIL) lines from improved early-flowering Senegalese cultivars and introduced germplasm, 10 F3 ISMP (ISRA Millet Population) lines from Senegalese and West African improved parents, 29 F4 PLS (Pearl Millet Landraces of Senegal) lines from late-flowering landraces, and 30 checks and other accessions. Checks represented contrasting phenotypic extremes: Gawane and Thialack 2 (positive/negative for flowering time), SL423 and Souna 3 (grain yield), and SL28 and IBV8004 (panicle length). The second and third sets were segregating populations advanced by pedigree selection with recurrent selfing, comprising 90 F1 genotypes (F1-Population) and 50 F2 genotypes (F2-Population), respectively, each derived from several parental crosses.

Field experimental design and phenotypic evaluation

Field work was conducted on experimental plots belonging to the Institut Sénégalais de Recherches Agricoles (ISRA) at the Centre National de Recherche Agronomique (CNRA), Bambey, Senegal. No field permits were required for this study, as all data collection was carried out on institutional research fields managed by ISRA, and access was granted by the institution. Trials were conducted between July and October 2024 across four locations in Senegal: Nioro du Rip (13°45′N, 15°48′W), Bambey (14.69°N, 16.45°W), Sinthiou Malem (13.77°N, 13.67°W), and Sefa (12.45°N, 15.32°W). The diversity panel was evaluated across all four locations using an alpha-lattice design (29 × 10) with three replications and a total surface area of 3302.85 m² per location. F₁-Population was tested at two locations, Nioro du Rip and Bambey (but only Bambey data are herein analyzed for reliability purposes), using an alpha-lattice design (9 × 10) with three replications and a total surface area of 931.5 m² per location. F₂-Population was evaluated across all four locations using a randomized complete block design (RCBD) with three replications and a total surface area of 2146.08 m² per location. Land preparation and agronomic management followed standard protocols across all sites to ensure uniform crop establishment and minimize environmental variability. Prior to sowing, fields were cleared, plowed, harrowed, and leveled to obtain a fine and uniform seedbed. Plots were laid out according to the respective experimental designs, and sowing was performed at a recommended spacing of 40 cm between hills and 75 cm between rows. Nitrogen Phosphorus Potassium (NPK-15-15-15) fertilizer was applied in two split doses, at planting and again 15 days after planting at a dose of 50 Kg/ha. Weed, pest, and disease management were carried out uniformly across all plots in accordance with local agronomic recommendations.

Data collection

Although multiple agronomic traits were recorded during the trials, the present study focused on four key traits: grain yield, panicle length, and plant height and flowering time. Data were collected at appropriate growth stages following standard practices to ensure accuracy and consistency across all sites. Flowering time was recorded as the number of days from sowing to 50% anthesis at the plot level, defined as the date on which at least half of the plants in a plot had begun shedding pollen from the primary spike. Grain yield was assessed at physiological maturity, once the crop had reached full grain fill and moisture content had stabilized; panicles were harvested plot-by-plot, threshed, and cleaned before weighing, with grain weight converted to kilograms per hectare using the effective harvested plot area and adjusted to a standard moisture content of 12.5% to allow valid comparisons across sites and harvest dates. Panicle length and plant height were both measured at physiological maturity on a subsample of three randomly selected plants per plot, with the mean value used as the plot-level observation. Panicle length was measured from the base of the primary panicle, immediately above the flag leaf node, to the tip of the terminal spikelet using a graduated ruler, with measurements restricted to primary panicles to avoid confounding effects of tillering. Plant height was recorded from ground level to the tip of the primary panicle, including the panicle.

Statistical analysis of phenotypic performance

A linear mixed model with no fixed effects beyond the intercept was fitted to estimate variance components, with all factors included as random effects, according to the following equation:

(1)

Where:

• y_ijkl is the observed phenotypic value,
• μ is the overall mean,
• (1|G_i) is the random effect of the i^th genotype,
• (1|Lj) is the random effect of the j^th location,
• (1|G × L) ij is the random effect of the genotype-by-location interaction,
• (1|Rk(j)) is the random effect of the k^th replication nested within the j^th location,
• (1|B l(kj)) is the random effect of the l^th block nested within the k^th replication and j^th location
• ε _ijkl is the residual error term, assumed to follow a normal distribution with mean zero and constant variance.

This full model was applied to the diversity panel, which was evaluated across the four locations, allowing estimation of all variance components including the genotype-by-location interaction. Because the F1-Population was evaluated at a single site and the F2-Population comprised segregating, non-fixed entries, variance components and broad-sense heritability were not estimated for these generations; they were summarized descriptively instead (see below). Broad-sense heritability for the diversity panel was calculated from the variance components using the following formula:

(2)

Where:

• = genetic variance (for genotypes)
• = genotype × location variance (where applicable)
• = residual variance
• = number of locations
• = number of replications per location

For GWAS, the Best linear Unbiased Estimators (BLUEs) for all the genotypes were used as phenotypic data using a linear mixed model where genotype was treated as fixed and location, block and replication as random. The model was written as follows:

(3)

Where:

• y_ijkl is the observed phenotypic value,
• μ is the overall mean,
• G_i is the fixed effect of the i^th genotype,
• (1|Lj) is the random effect of the j^th location,
• (1|G × L) ij is the random effect of the genotype-by-location interaction,
• (1|Rk(j)) is the random effect of the k^th replication nested within the j^th location,
• (1|B l(kj)) is the random effect of the l^th block nested within the k^th replication and j^th location
• ε _ijkl is the residual error term, assumed to follow a normal distribution with mean zero and constant variance.

The coefficient of variation among genotypes was calculated as the standard deviation of genotype BLUEs divided by their mean, and was used to assess phenotypic variability across the panel. This statistic was used to assess the relative magnitude of phenotypic variability among genotypes in the GWAS panel independently of trait scale, and to flag traits with high residual variability that may affect association power. Pearson’s correlation was also computed between the four studied traits in the GWAS panel. All statistical analyses mentioned above were conducted using the R software version 4.5.1 primarily employing the lme4 package for mixed model analysis.

DNA sample collection

Leaf tissue for genotyping was sampled from 279 individuals, combining 178 entries drawn from the 290-genotype diversity panel combined with 53 F1 and 48 F2 individuals from the two segregating populations. Genetic-diversity metrics were therefore computed on the overall 279-genotyped collection, while genome-wide association analysis was restricted to the 137 most advanced, genetically stable entries (PMiGAP inbred lines, F5 ISMI lines, and checks) selected from the genotyped material. Samples were collected using leaf punchers and placed directly into 96-well collection plates. To prevent cross-contamination between genotypes, punchers were systematically sterilized with 70% ethanol prior to each successive sampling. Collected samples were then shipped to SEQART AFRICA, hosted at the International Livestock Research Institute (ILRI) in Nairobi, Kenya, for genotyping. DNA extraction was performed from leaf tissue using the NucleoMag Plant DNA Extraction Kit following the manufacturer’s protocol. The quantity and quality of extracted genomic DNA were assessed by gel electrophoresis on a 0.8% agarose gel, and DNA concentrations were confirmed to fall within the range of 50–100 ng/µl, meeting the quality thresholds required for downstream genotyping applications.

Genotyping method

Genotyping was performed using the DArTseq Genotyping-by-Sequencing (GBS) platform, which provides rapid and high-quality genome-wide profiling. Library preparation followed the protocol described by Kilian et al. (2012) [22]. Briefly, complexity reduction was achieved through double digestion of genomic DNA using a combination of PstI and HpaII restriction enzymes, followed by ligation of barcoded adapters and a common adapter. Adapter-ligated fragments were then amplified by PCR to generate the final sequencing libraries. Sequencing was carried out on an Illumina NovaSeq X instrument using single-read runs of 138 cycles, generating high-throughput sequence data for downstream SNP discovery and genotype calling.

SNP calling and filtering criteria

Raw marker scoring was performed using DArTsoft14, an in-house proprietary pipeline that applies sequence-based algorithms to score two types of markers: SilicoDArT markers and SNP markers. Both marker types were scored in binary format, representing the presence (1) or absence (0) of a restriction fragment carrying the marker sequence in the genomic representation of each sample. All markers were aligned to the pearl millet reference genome assembly PearlMillet.Tift [23] to enable chromosomal anchoring and positional annotation. Following alignment, a quality control filtering pipeline was applied to retain only high-confidence SNP markers for downstream analysis. Markers were filtered based on three criteria: a minor allele frequency (MAF) greater than 5%, a missing data rate per sample below 20%, and a missing data rate per SNP below 10%. SNPs failing any of these thresholds were excluded, resulting in a high-quality filtered SNP dataset used for all subsequent genomic analyses.

Genetic diversity metrics

Genetic diversity metrics were computed for the full collection of 279 individuals using the snpReady package in R. For each SNP, the following statistics were estimated: minor allele frequency (MAF), expected heterozygosity (gene diversity, DG), observed heterozygosity (HO), polymorphism information content (PIC), inbreeding coefficient (F), and selection coefficient (S). Summary statistics were averaged across all retained polymorphic loci and reported by breeding source: Group 1 (international PMiGAP lines) and Group 2 (local breeding material, including check cultivars and breeding lines at various generations), as well as for the complete panel.

Population structure analysis

Population structure was examined only within the GWAS panel made of 137 genotypes through principal component analysis and partitioning-around-medoids (PAM) [24] clustering of Euclidean distances computed from standardized SNP dosages using the Cluster R package. This method partitions individuals into k clusters by identifying the most representative central object, or medoid, within each cluster. The optimal number of genetic clusters (k) was determined by testing k values ranging from 2 to 8 and selecting the value yielding the highest average silhouette width. To visualize population structure, principal component analysis (PCA) was applied to the filtered genotype matrix, and a three-dimensional PCA plot was generated with PAM cluster assignments overlaid using the fviz_cluster () function from the factoextra R package. Because structure was modelled through these continuous PCA + K terms rather than discrete cluster membership, the PAM partition is descriptive only and does not enter the GWAS model.

Linkage disequilibrium analysis

Linkage disequilibrium (LD) decay was also assessed across the GWAS panel to characterize the extent of allelic associations among SNP markers and to inform the resolution of genome-wide association analyses. Prior to LD computation, the VCF file was converted to Genomic Data Structure (GDS) format using the snpgdsVCF2GDS function of the SNPRelate package [25] in R program. Pairwise LD was estimated chromosome-by-chromosome using the snpgdsLDMat function, which computes the Pearson correlation coefficient (r) between allele dosages at pairs of SNP loci within a sliding window of 500 consecutive markers. The squared correlation coefficient (r²) was derived and used as the primary measure of LD. Only SNP pairs separated by a physical distance of up to 500 kb were retained for downstream analysis. Pairwise r² values were subsequently binned into non-overlapping 5 kb distance classes, and the mean r² was computed within each bin to characterize the genome-wide decay trend. The physical distance at which mean r² dropped below the conventional threshold of 0.1 was used as the operational estimate of LD extent, following the recommendation of Flint-Garcia et al. (2003) [26].

Genome-wide association analysis

Association analysis was conducted in a panel consisting of the most advanced and stable genotypes (F₅ ISMI lines, PMiGap inbred lines and local checks). Association Analyses were performed using the mrMLM (Multi-locus Random-SNP-Effect Mixed Linear Model) [27] package version 4.0.2. This method integrates six multi-locus models to enhance QTL detection while controlling for false positives: (i) mrMLM, (ii) FASTmrMLM, (iii) FASTmrEMMA, (iv) pKWmEB, (v) pLARmEB, and (vi) ISIS EM-BLASSO. Population structure (PCA) and kinship (K) matrices were incorporated into all models to correct for population stratification and individual relatedness. Although mrMLM applies a default critical LOD of 3.0, we adopted a more stringent genome-wide threshold of LOD 6.5 to limit false positives. This corresponds to a Bonferroni-adjusted significance level of α ≈ 0.005 across the 14,718 markers (−log₁₀(0.005/14718) ≈ 6.5). The general model used in the multi-locus mixed model framework is described in the equation below:

(4)

Where:

• Y is the phenotypic trait value,
• βo is the intercept of the model,
• β_i is the effect size (regression coefficient) of the ith SNP genotype (coded as 0,1 or 2 for the number of minor alleles),
• X_i is the genotype score for ith SNP,
• ϕ_j is the effect size of the jth covariate,
• Z_j is the value of the jth covariate for each individual (population structure),
• ɛ is the error term associated with all other unmeasured factors influencing the trait.

Candidate gene analysis

To identify putative candidate genes underlying the detected MTAs, a physical window based on LD result upstream and downstream of each significant SNP was defined using the pearl millet reference genome (Tift23D2B1-P1-P5). All annotated genes located within this window were extracted from the corresponding genome annotation file (GFF3 format) [23] and considered as candidate genes for the corresponding marker-trait association. All analyses were conducted in R version 4.5.1 using dedicated Bioconductor packages, including GenomicRanges, rtracklayer, Biostrings, and GenomicFeatures. Genes identified within the defined search window were retained for downstream functional annotation to further characterize their putative roles in the regulation of the traits under study.

Analysis of variance in the diversity panel

Analysis of variance in the diversity panel revealed significant differences (p < 0.001) among genotypes for all four traits across all three sets (Table 1). Location effects were highly significant for yield, panicle length, and flowering time, but non-significant for plant height. Genotype-by-location interaction was significant for all traits (p < 0.01 to p < 0.001). Broad-sense heritability estimates ranged from 0.75 to 0.91 for grain yield and panicle length respectively.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Phenotypic variation of the studied traits in the diversity panel.

https://doi.org/10.1371/journal.pone.0358045.t001

Descriptive statistics in the early segregating material

Descriptive statistics for the four agronomic traits in the F1 and F2 generations are summarized in (Table 2). Flowering time was the least variable trait in both generations, averaging 58.7 days in the F1 (range 47–69) and 56.7 days in the F2, where it spread more widely (range 41–74). Plant height averaged 229.1 cm in the F1 (range 158–282) and 218.3 cm in the F2, with a comparable spread but a markedly wider range (133–308 cm). Panicle length followed a similar pattern, declining from 40.1 cm in the F1 (range 22–65) to 37.3 cm in the F2, where variability increased appreciably (range 20–71). Grain yield was the most variable trait: it averaged 687.0 kg ha^-1 in the F1 and 462.1 kg ha^-1 in the F2, but in both generations the standard deviation approached or exceeded the mean (774.0 and 443.3, respectively), and values ranged from around 100 to over 4400 kg ha^-1. Across all the four traits, the F2 consistently spanned a wider range than the F1, consistent with the transgressive segregation expected in early-generation breeding material.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Descriptive statistics for the four agronomic traits in the F1 and F2 generations.

https://doi.org/10.1371/journal.pone.0358045.t002

Phenotypic performance and variability in the GWAS panel

The frequency distributions of the four agronomic traits in the GWAS panel are presented in (Fig 1), and the corresponding descriptive statistics are summarized in (Table 3). Panicle length (Fig 1A) displayed a roughly bell-shaped distribution centered around 38–42 cm, spanning from approximately 18–63 cm, with a CV of 23.60%. Plant height (Fig 1B) showed a broad, relatively uniform distribution ranging from approximately 150–270 cm, with a concentration of genotypes in the 200–250 cm range and a CV of 12.90%. Flowering time (Fig 1C) followed an approximately normal distribution, with values ranging from around 40–72 days, the majority of genotypes flowering between 50 and 62 days, and a CV of 8.66%. Grain yield (Fig 1D) exhibited a right-skewed distribution ranging from below 500 to nearly 4000 kg/ha, with the majority of genotypes concentrated between 1000 and 2500 kg/ha and a long tail toward higher yield values, and a CV of 43.56%.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Descriptive statistics and coefficient of variation (CV) for agronomic traits in the GWAS panel.

https://doi.org/10.1371/journal.pone.0358045.t003

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Distribution of the four studied traits across the GWAS panel.

https://doi.org/10.1371/journal.pone.0358045.g001

Trait correlations

Among the studied traits there were varying degrees of correlation (Fig 2). Plant height showed a moderate positive correlation with panicle length (r = 0.52, p ≤ 0.001) and grain yield (r = 0.45, p ≤ 0.001). Similarly, panicle length was positively correlated with yield (r = 0.28, p ≤ 0.001). In contrast, flowering time displayed weak correlations with other traits. It showed a slight positive correlation with plant height (r = 0.08, p ≤ 0.001) and panicle length (r = 0.16, p ≤ 0.001), but a moderate negative correlation with yield (r = –0.28, p ≤ 0.001).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Correlation coefficients among the studied traits.

Legend for Significance Stars: *** p ≤ 0.001; ** p ≤ 0.01.

https://doi.org/10.1371/journal.pone.0358045.g002

Marker coverage and distribution

The distribution of the 14718 refined SNP markers was plotted in 1 Mb (megabase pair) window across the P. glaucum genome (Fig 3). Variant distribution was fairly uniform across the chromosomes. The highest density was observed in the telomeric region of Chromosome 6. The total number of SNPs per chromosome ranged from 1866 SNPs on chromosome 6–2385 SNPs on chromosome 2.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Distribution of SNPs on the 7 chromosomes of pearl millet within 1Mb window size.

The horizontal axis shows chromosome length with their respective SNPs density. The legend (0- > 64) indicates the SNPs density.

https://doi.org/10.1371/journal.pone.0358045.g003

Genetic diversity metrics

Genetic diversity metrics were broadly similar across the two breeding groups and the complete collection (Table 4). Group 1 (PMiGAP) showed slightly lower values for gene diversity (DG = 0.27), PIC (0.22), and inbreeding coefficient (F = 0.29) compared to Group 2 (DG = 0.30, PIC = 0.24, F = 0.32), while MAF was nearly identical (0.20 vs 0.21). Observed heterozygosity was identical across all three groups (HO = 0.20), and the selection coefficient was slightly higher in Group 2 (S = 0.40) than in Group 1 (S = 0.38). Overall, the complete collection metrics closely mirrored those of Group 2, reflecting the larger contribution of local breeding material to the full dataset.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Genetic diversity metrics estimated in the complete genotyped panel (n = 279). Group 1 (PMiGAP panel) and Group 2 (local breeding material).

https://doi.org/10.1371/journal.pone.0358045.t004

Population structure analysis in the complete panel

PAM clustering identified a two-group partition as the optimal solution (average silhouette width maximal at k = 2), comprising 44 and 93 entries (Fig 4). Differentiation between the two groups was nonetheless weak: the first three principal components jointly captured less than 9% of the genotypic variance, average silhouette widths remained below 0.05. The two clusters showed some overlap in principal component space. The larger group was dominated by ICRISAT-derived (PMiGAP) lines, whereas the smaller group was enriched for locally developed ISMI and Senegalese check materials, reflecting a loose tendency to group by breeding origin. Because the association model corrected for structure and relatedness through continuous PCA + kinship terms rather than cluster membership, this partition is descriptive only and had no bearing on the detected MTAs.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Population structure in the 137-entry GWAS panel.

Left: principal component analysis of the 137 advanced entries, colored by partitioning-around-medoids (PAM) cluster at the silhouette-optimal k = 2 (Cluster 1, n = 44; Cluster 2, n = 93); the first three components explained 4.1%, 2.4%, and 1.9% of the genotypic variance and the clusters overlapped extensively. Right: average silhouette width peaked at k = 2 (dashed line) but stayed below 0.05 for all k, indicating no strongly differentiated subpopulations within the panel.

https://doi.org/10.1371/journal.pone.0358045.g004

Linkage disequilibrium analysis

Linkage disequilibrium (LD) decay was assessed across the GWAS panel by computing mean r² values between all pairwise SNP combinations as a function of physical distance (kb). LD declined rapidly with increasing genomic distance, dropping from r² ≈ 0.10 at near-zero distance to the conventional threshold of r² = 0.1 at approximately 8 kb (Fig 5). Beyond this threshold, LD continued to decline progressively, stabilizing at a residual background level of r² ≈ 0.02–0.04 beyond approximately 50 kb. The locally estimated scatterplot smoothing (LOESS) curve confirmed the overall decay trend and was consistent with the raw mean r² profile.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Linkage disequilibrium (LD) decay across the GWAS panel.

Mean r² values (blue line) are plotted against physical distance (kb) between SNP pairs. The locally estimated scatterplot smoothing (LOESS) curve (red line) highlights the overall decay trend. LD decayed below the threshold of r² = 0.1 (dashed grey line) at approximately 8 kb.

https://doi.org/10.1371/journal.pone.0358045.g005

Marker effects on agro-morphological traits

Multi-locus GWAS identified a total of 11 significant marker-trait associations (MTAs) distributed across 6 of the 7 pearl millet chromosomes, with LOD scores ranging from 6.83 to 10.78 (Table 5). Grain yield was associated with three MTAs on chromosomes 2, 4, and 5, with R² values ranging from 6.18% to 48.67%. Plant height showed the highest number of associated loci, with five MTAs detected on chromosomes 3, 4, 5, 6, and 7, explaining between 8.07% and 19.0% of phenotypic variance. Panicle length was associated with two MTAs on chromosomes 2 and 4, with R² values of 10.52% and 13.83%. Flowering time was associated with a single MTA on chromosome 2 (SNPPM2516, LOD = 8.28, R² = 5.92%). Allele effects were variable in direction across traits and loci, ranging from −660.24 to 208.25. The Manhattan and Q-Q plots for all traits are presented in (Fig 6).

Download:

• PNG
  larger image
• TIFF
  original image

Table 5. Marker-trait associations (MTAs) detected in chromosomal regions using multi-locus genome-wide association analysis.

https://doi.org/10.1371/journal.pone.0358045.t005

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Manhattan and quantile-quantile (Q-Q) plots for the four agronomic traits.

Results are shown for grain yield (A), plant height (B), panicle length (C), and flowering time (D). For each trait, the Manhattan plot (left) displays the − log₁₀(p-value) for each marker across the seven chromosomes of pearl millet, with the secondary axis (pink) indicating the LOD score. The horizontal grey line represents the LOD threshold of 6.5 applied for declaring significant associations. Markers in red were identified by more than one model.

https://doi.org/10.1371/journal.pone.0358045.g006

Candidate gene analysis

To identify putative candidate genes underlying the significant MTAs, significant SNPs were annotated against the reference genome and overlapping genes were retrieved (Table 6); functional annotations were inferred by homology and, where indicated, transferred from related species such as Setaria italica, and should therefore be read as positional hypotheses rather than demonstrated causal genes. For grain yield, associations on chromosome 4 co-localized with two receptor-like kinase genes, a cysteine-rich receptor-like protein kinase 6 (dpca4g263180) and a DUF26/TKL_IRAK kinase (dpca4g263190), whose families have been implicated in stress perception and signal transduction, while on chromosome 5 significant markers overlapped an expansin-A12 gene (dpca5g385240), annotated as a cell-wall-loosening protein linked to cell expansion and grain development, and a fatty acid–CoA ligase (dpca5g385250) involved in lipid metabolism.

Download:

• PNG
  larger image
• TIFF
  original image

Table 6. Candidate genes directly overlapping significant SNPs identified in the pearl millet panel. Gene positions are based on the reference genome annotation, and only SNPs located within 8 kb window annotated gene regions are presented.

https://doi.org/10.1371/journal.pone.0358045.t006

For plant height, candidate genes included a hydroxyproline-rich glycoprotein (dpca6g429350) and an RWP-RK/RKD5-like transcription factor (dpca6g429360) on chromosome 6, annotated with roles in cell-wall structure and developmental regulation, together with an isocitrate lyase (dpca7g479100) on chromosome 7 and an additional chromosome 4 gene of unknown function (dpca4g304390).

For panicle length, a notable candidate was an AP2/ERF transcription factor on chromosome 4 (dpca4g323040), a family with established roles in inflorescence and organ development, whereas a second chromosome 4 locus corresponded to a transposable-element-derived integrase domain (dpca4g323050) and a chromosome 2 marker overlapped an aspartic-protease/xylanase-inhibitor gene (dpca2g122560).

For flowering time, a single significant association on chromosome 2 co-localized with an indole-3-acetate β-glucosyltransferase (dpca2g096240), a gene annotated in auxin homeostasis during developmental transitions. Chromosome 4 thus carried candidate genes for grain yield, plant height, and panicle length; as noted above, however, these lie at widely separated positions and represent physically independent loci rather than a single pleiotropic region. Among the annotated candidates, the developmental genes expansin-A12, the AP2/ERF transcription factor, and the auxin-conjugating β-glucosyltransferase offer the most plausible functional hypotheses for the associated agronomic traits and are accordingly prioritized for validation.

Phenotypic performance analysis

Analysis of variance revealed highly significant genotypic effects for all four studied traits in the diversity panel, confirming the presence of exploitable genetic variation in the evaluated pearl millet panel. Similar significant genotypic differences for these traits have been reported in previous pearl millet studies conducted across West African environments [9,28]. The significant genotype × location (G × L) interaction detected for all traits underlines the sensitivity of these traits to environmental variation, a finding consistent with reported multi-environment analyses of pearl millet and sorghum agromorphological performances [29].

The levels of phenotypic variability observed in this study are consistent with those reported in pearl millet diversity panels evaluated across West Africa and the broader semi-arid tropics [12,30], confirming that the Senegalese germplasm captures a representative range of genetic diversity for the traits studied. The high CV for grain yield reflects the phenotypic heterogeneity of the panel entries rather than residual segregation; this heterogeneity is expected given that the panel draws on entries from different breeding sources. The relatively low variation in flowering time compared to yield-related traits follows a pattern widely documented in pearl millet, where photoperiod sensitivity and adaptation to terminal drought impose strong directional selection on phenology while leaving morphological and yield traits more open to diversification [31].

Heritability estimates across traits were consistent with the quantitative genetic architecture expected for pearl millet under Sahelian conditions. The relatively high H² for grain yield indicates a substantial genetic basis for yield variation in this panel, comparable to values reported from combining ability and multi-environment studies in West African pearl millet [32], though the residual environmental component underscores the importance of multi-environment evaluation for reliable selection gains. The high heritability observed for panicle length and plant height reflects their predominantly additive genetic architecture and lower sensitivity to G × E interaction, consistent with findings across recombinant inbred line populations and diversity panels in pearl millet [33,34]. The moderate to high heritability for flowering time aligns with reports from Senegalese landraces [35], where photoperiod sensitivity introduces trial-to-trial variation while maintaining a meaningful genetic signal. These estimates suggest that direct phenotypic selection will be most efficient for morphological traits, while grain yield improvement will require multi-environment testing and genomic-assisted approaches to effectively capture and exploit the available genetic variation.

Phenotypic distribution and trait correlation

The phenotypic distribution across the GWAS panel revealed substantial variation for all four traits, consistent with quantitative inheritance and phenotypic ranges reported in pearl millet panels evaluated across West Africa and the semi-arid tropics [12,30]. The correlation structure reflected biologically meaningful associations consistent with patterns widely documented in pearl millet: the positive relationships among plant height, panicle length, and grain yield suggest shared developmental pathways and pleiotropic genetic control, while the negative correlation between flowering time and grain yield reflects the drought escape mechanism whereby early flowering confers yield stability under terminal drought, a priority selection criterion in Sahelian breeding programs [31]. These patterns support the use of indirect selection through morphological traits under favorable conditions and confirm early flowering as a key target for adaptation to the Senegalese rainfall pattern.

Genetic diversity metrics, population structure

The genetic-diversity indices and the population-structure analysis were computed on complementary subsets rather than on a single panel, and are best read together. Gene diversity, PIC, observed heterozygosity, and inbreeding coefficients were estimated on the full genotyped set (279 genotypes), which still included early-generation segregating material, whereas the PAM/PCA structure analysis was restricted to the 137 advanced, genetically stable entries that formed the association panel. The moderate gene diversity (GD) and PIC observed in both groups are consistent with previous SNP-based characterizations of pearl millet germplasm, where these indices typically range from 0.20 to 0.40 depending on origin, breeding history, and marker density [12,36]. The slightly lower diversity in Group 1 (PMiGAP inbreds) relative to Group 2 reflects the cumulative effect of directional selection within the ICRISAT breeding program, since elite pools under sustained improvement predictably harbour narrower allelic ranges than locally adapted or multi-cross material [37]. Group 2, by contrast, is compositionally heterogeneous: check cultivars alongside multi-cross lines spanning early to advanced generations, and this breadth sustains higher allelic diversity through the accumulated recombination and residual segregation characteristic of early-generation material, accounting for its elevated GD and PIC. These contrasts bear directly on how heterozygosity was interpreted and how the association panel was assembled. The near-identical observed heterozygosity across the two groups is consistent with the allogamous, protogynous reproductive biology of pearl millet, and the moderate inbreeding coefficients confirm that the genotyped diversity set had not reached fixation. Because non-fixed genotypes violate the fixed-line assumption underlying single-marker association testing and can inflate association noise, the early-generation segregating material was excluded and the GWAS was restricted to the 137 stable entries; residual relatedness within that panel was then controlled through a kinship matrix in the mixed-model framework [38,39]. Finally, although the PIC values fall at the lower end of the moderately informative range [40], they remain adequate for association analysis given the large number of markers retained after quality filtering, while the minor-allele-frequency spectrum spans rare variants of potential relevance to adaptation in the Sahelian environment.

Population structure within the 137-entry association panel was weak and largely continuous. PAM clustering identified k = 2 as the silhouette optimum, yet average silhouette widths remained below 0.05 across all values of k, indicating that the panel does not resolve into sharply differentiated subpopulations. This near-continuous stratification is consistent with the panel’s composition, in which PMiGAP lines and local multi-cross material derived from ICRISAT parents share substantial ancestry rather than forming discrete pools. Soft genetic boundaries of this kind are expected in an obligate outcrosser, where protogynous flowering and wind pollination sustain continuous gene flow that prevents sharp discontinuities [41,42]. The low variance captured by individual principal components is likewise a known property of high-dimensional differentiation in outcrossing cereals rather than a shortcoming of the panel, and has been reported for comparable West African germplasm [36,37]

Methodologically, continuous stratification of this form favours fitting several leading PCs as fixed covariates together with a kinship matrix as a random effect, rather than a single Q-matrix, to control false positives while preserving detection power [43]. The panel therefore sits at a productive point in the diversity–stratification trade-off: its allelic range is broad enough to power association detection, while its structure is correctable rather than confounding. Association signals recovered consistently across the panel are expected to be the most transferable across the ISRA–ICRISAT breeding continuum. Finally, the faster LD decay characteristic of West African pearl millet relative to South Asian material [12,36] means associations from Indian panels do not transfer directly to this context, reinforcing the value of a regionally anchored reference panel for future multi-trait and multi-environment GWAS in the Sahelian production system.

Linkage disequilibrium analysis

Linkage disequilibrium decayed rapidly, with mean r² falling below 0.1 at approximately 8 kb across the 14,718 SNPs scored in the GWAS panel, consistent with the outcrossing biology of pearl millet, in which recurrent recombination erodes marker–marker correlations over short distances. The reference PMiGAP diversity panel reported even faster decay (r² < 0.2 within 84–444 bp), a rate comparable to allogamous maize [11]. The greater LD extent observed here is expected from the panel’s composition: unlike a largely panmictic diversity set, our panel combined advanced inbred lines with locally derived breeding material, several entries sharing pedigree and subject to directional selection, both of which reduce effective recombination and extend haplotype blocks. Such subpopulation-dependent variation is documented in pearl millet, where the West-African subgroup shows greater decay owing to a longer recombination history [36]. The short-range LD confirms that dense, genome-wide marker coverage is required to capture causal variants [44].

Marker effects on agro-morphological traits and candidate gene analysis

Eleven MTAs were detected across six chromosomes for the traits under study, each association individually explaining a modest fraction of the overall phenotypic variance. This distribution is characteristic of the polygenic, small-effect architecture typical of quantitative agronomic traits in pearl millet, and it aligns with prior genome-wide work in the crop: the reference GWAS across test-cross progenies identified 1,054 significant MTAs for fifteen traits, with individual associations explaining between 9% and 27% of phenotypic variation [11]. The only exception in our study was on chromosome 4, where the marker SNPPM6939 associated with grain yield explained 48.67% of the phenotypic variance and had a negative effect on yield. An effect of this magnitude is atypical for a quantitative trait and should be interpreted with caution: in a panel of only 137 entries with no independent validation population, single-locus effect sizes are prone to substantial overestimation (the Beavis effect), and the value may partly reflect minor-allele frequency or allele-class imbalance rather than a true major effect. This need for independent confirmation is reinforced by the deleterious-variant burden documented in Senegalese pearl millet [45], which argues for building association resources on regionally adapted germplasm and validating candidate signals before they are deployed in selection.

Several of other associations correspond, at the chromosome level, to genomic regions previously implicated in the same traits. Earlier marker-based studies in pearl millet associated marker alleles with plant height on linkage group 7, and with panicle length on linkage groups 2, 5, and 7, while a further allele on linkage group 7 was associated with grain yield [44], and more recent association analyses have likewise recovered agronomic-trait associations for plant height, panicle length, and grain weight in the crop [46]. This convergence of independent panels and marker systems on broadly overlapping regions is consistent with these being genuine effects rather than artefacts; the correspondence is, however, at the chromosome level and does not by itself establish that the same underlying loci are involved.

A notable feature of the present results is that chromosome 4 harbored associations for three traits: grain yield (SNPPM6939, 3.9 Mb), plant height (SNPPM8294, 128.2 Mb), and panicle length (SNPPM8661, 227.9 Mb), making it, together with chromosome 2, the most trait-rich region in the panel. This co-occurrence does not, however, imply a shared genetic basis. The three markers are dispersed across more than 220 Mb of the chromosome, with the plant-height and panicle-length signals alone separated by roughly 100 Mb. Because linkage disequilibrium in this panel decayed to r² < 0.1 within approximately 8 kb, associations this far apart cannot be in linkage with one another and instead represent physically independent loci that merely reside on the same chromosome. The co-localization of plant height and panicle-related traits reported previously in pearl millet [47] therefore does not extend to the present associations, which show no evidence of a single pleiotropic region governing multiple traits. The functional candidates underlying each of these independent signals are considered below. Inspection of the ± 8 kb intervals flanking the significant markers nominated several functionally credible candidates, which should be regarded as hypotheses for subsequent validation rather than confirmed causal genes. Because these candidates were identified solely by physical proximity to significant markers within LD-defined windows, their positions localize but do not establish a causal relationship, and the links proposed below are positional hypotheses to be tested functionally. Two gene families recurred in the intervals flanking the plant-architecture and panicle associations: F-box domain proteins and AP2/ERF transcription factors. The AP2/ERF family is a well-characterized regulator of inflorescence development and yield in cereals: in rice, these transcription factors participate in the regulation of panicle architecture, grain plumpness, and yield formation by controlling inflorescence meristem activity, cell division, and differentiation [48], and their agronomic relevance extends across grasses, in wheat, mutation of an AP2/ERF gene (DUO-B1) regulating spike architecture increased grain number per spike and raised field yield without penalizing other major agronomic traits [49]. The presence of an AP2/ERF candidate near the chromosome 4 region jointly associated with plant height and panicle length would therefore be consistent with a role in meristem determinacy and inflorescence patterning, though this remains to be demonstrated. The F-box candidates are similarly noteworthy: in rice, mutations in the F-box gene LARGER PANICLE improve panicle architecture and enhance grain yield [50], and non-synonymous variants in both an AP2/ERF gene and an F-box domain-containing protein were recently reported to jointly explain a large proportion of the variance for panicle exsertion [51]. The recurrence of these two families in our intervals parallels that functional pairing and supports their prioritization as candidates for follow-up, rather than confirming their involvement in the panicle-related associations.

For flowering time, the strongest a priori candidate in pearl millet is the phytochrome pathway. Natural variation at PHYTOCHROME C is a long-recognized determinant of phenology and morphology in the crop: association studies first linked natural variation at PHYC to flowering time and morphological variation across a set of inbred lines [41], a result subsequently refined through fine-scale association mapping and analysis of linkage disequilibrium and selection in the vicinity of the PHYTOCHROME C gene [52]. This locus has since been repeatedly implicated in Sahelian adaptation, having been identified as a major flowering-time gene underpinning selection for earlier flowering in response to climatic variation [31]; it is cited here as regional context, as the flowering association recovered in the present panel does not fall within the PHYC region. Flowering time is highly quantitative [53], so the single association recovered here most likely reflects the limited power of a panel of this size under a stringent multi-locus threshold, rather than a truly simple genetic basis, with a larger panel expected to resolve more loci. It is nonetheless tempting to speculate that crosstalk between photoreceptor and auxin-mediated pathways contributes to the phenological variation observed here, a hypothesis consistent with documented phytochrome–auxin interactions in other species [54] but one that would require direct functional testing to validate.